RTP Payload Format for the iSAC Codec

This document gives a general description of the iSAC wideband speech codec and specifies the iSAC payload format for usage in RTP packets. Also included here are the necessary details for the use of iSAC with the Session Description Protocol (SDP).

The iSAC codec is an adaptive wideband/superwideband speech and audio codec that operates with short delay, making it suitable for high quality real time communication. It is specially designed to deliver wideband speech quality in both low and medium bit rate applications. It also handles non-speech audio well, such as music and background noise . The iSAC codec compresses speech frames of 16 kHz, 16-bit sampled input speech, each frame containing 30 or 60 ms of speech. It also has a superwideband mode which allows a 32 kHz sampling rate. In super-wideband mode the input signal is split into wideband (0-8 kHz) and upper (8-16 kHz) signal. Each sub-band is encoded independently, and their associated payloads concatenated, c.f. Figure 2, to construct the overall iSAC super-wideband RTP payload. Note that the same encoder/decoder is used for the wideband part for both wideband and super-wideband modes. The codec runs in one of two different modes called channel-adaptive mode and channel-independent mode. In both modes iSAC is aiming at a target bit rate, which is neither the average nor the maximum bit rate that will be reach by iSAC, but corresponds to the average bit rate during peaks in speech activity. The bit rate will sometimes exceed the target bit rate, but most of the time will be below. The average bit rate obtained is on average about a factor of 1.2 times lower than the target bit rate on continuous speech, and will be lower on speech with pauses. In channel-adaptive mode the target bit rate is adapted to give a bit rate corresponding to the available bandwidth on the channel. The available bandwidth is continuously estimated at the receiving iSAC and signaled in-band in the iSAC bit stream. Even at dial-up modem data rates (including IP, UDP, and RTP overhead) iSAC delivers high quality by automatically adjusting transmission rates to give the best possible listening experience over the available bandwidth. The default initial target bit rate is 20000 bits per second in channel-adaptive mode. In channel-independent mode a target bit rate has to be provided to iSAC prior to encoding; the target bit rate can be changed over the time of the call. After encoding the speech signal the iSAC coder uses lossless coding to further reduce the size of each packet, and hence the total bit rate used. The adaptation and the lossless coding described above both result in a variation of packet size, depending both of the nature of speech and the available bandwidth. Therefore, the iSAC codec, in wideband mode, operates at transmission rates from about 10 kbps to about 32 kbps. In super-wideband mode, the transmission rate is in the range of 10 kbps to 56 kbps. If operating in super-wideband mode, the iSAC codec automatically adjusts the effective encoded audio bandwidth for the best experience.

The main characteristics can be summarized as follows: Wideband or superwideband, 16 kHz or 32 kHz respectively, speech and audio codec Variable bit rate, which depends on the input signal Adaptive rate with two modes: channel-adaptive or channel-independent mode Bit rate range from around 10 kbps to 32 kbps when operating on wideband input. For input audio sampled at 32 kHz, the bit rate range 10 kbps to 56 kbps. Operates on 30 or 60 ms of speech for wideband inputs, and only 30 ms for super-wideband inputs. In super-wideband mode, depending on the target bit rate, the effective bandwidth is adjusted for the optimal experience.

The iSAC codec in wideband mode uses a sampling rate clock of 16 kHz, so the RTP timestamp MUST be in units of 1/16000 of a second. In super-wideband mode, the iSAC codec uses a sampling rate clock of 32 kHz, so the RTP timestamp MUST be in units of 1/32000 of a second. The RTP payload for iSAC has the format shown in Figure 1. No additional header fields specific to this payload format are required. For RTP based transportation of iSAC encoded audio, the standard RTP header is followed by one payload data block.

Figure 1: RTP packet format for iSAC

The iSAC payload block consists of a payload header and one or two encoded 30 ms speech frames. The iSAC payload is generated in the following manner: Parameters representing one or two 30 ms frames of speech data are determined by the encoder. The parameters are quantized to generate encoded data corresponding to the one or two speech frames. The length of the encoded data is variable and depends on the signal characteristics and the target bit rate. The payload header is generated (described in Section 3.2) and added before the encoded parameter data for the speech frame(s). Lossless coding is applied to the complete iSAC payload block, including payload header, to generate a compressed payload. The length depends on the length of the data generated to represent the speech and the effectiveness of the lossless coding. No part of the payload header or the encoded speech data can be retrieved without partly or fully decoding the packet. The following figure shows an iSAC payload block containing 60 ms of encoded speech data.

Figure 2: Payload format for iSAC

The payload header holds information for the receiver about the available bandwidth, in the form of a Bandwidth Estimation Index (BEI), and the length of the speech data in the current payload (frame length, FL). The header has the format defined in Figure 3. Note that the size of the header can vary due to the lossless encoding described in section 2 and in section 3.1. Also note that the BEI is always estimated and transmitted, even if iSAC runs in channel-independent mode.

Figure 3: Payload Header BEI: Bandwidth Estimation Index. The bandwidth estimate is quantized into one out of 24 values. Valid values are 0 to 23. FL: The length of the speech data (Frame Length) present in the payload, given in number of speech samples. Valid frame lengths are 480 (30 ms) and 960 (60 ms) samples.

The iSAC encoded speech data consist of parameters representing one or two frames of 30 ms speech. The length of the speech data is signaled in the header (in number of samples), and the length may change at any time during a session. In channel-adaptive mode the length is changed to best utilize the available bandwidth, and extra padding is added to some packets as a bandwidth probe. The iSAC payload is padded to whole octets, and has a variable length depending on the input source signal, number of 30 ms speech frames, and target bit rate. The number of octets used to describe one frame of 30 ms speech typically varies from around 50 to around 120 octets. For the case of 60 ms speech (two 30 ms speech frames), the number of octets varies from around 100 to around 240 octets. The absolute maximum allowed payload length is 400 octets. The user can choose to lower the maximum allowed payload length. Minimum value is 100 octets. It is possible for the user to choose a maximum bit rate (averaged over a frame) instead of a maximum payload length. The maximum payload length is then dependent on the length of the speech data represented in the payload (30 or 60 ms). Possible maximum rates are in the range of 32000 to 53400 bits per second. The sensitivity to bit errors is equal for all bits in the payload.

In super-wideband mode, payloads associated with each sub-band (wideband 0-8 kHz and upper-band 8-16 kHz) are constructed independently and concatenated as depicted in Figure 2. Note that in super-wideband mode only one 30 ms frame is encoded in each payload.

| ]]> Figure 4: Super-Wideband payload format Payloads of wideband and upper-band are encoded independently, allowing the encoder to simply concatenate two payloads to construct one iSAC super-wideband payload. The RTP payload of the iSAC super-wideband codec starts with the payload of the wideband part, which is padded to whole octets, followed by one byte (LEN in Figure 4) representing the length of the remaining sequence, payload of the upper-band plus 4 bytes for CRC sequence. If LEN_UB denotes the length of the upper-band payload, then LEN = 1 + LEN_UB + 4. This value should not exceed 255, otherwise upper-band payload is omitted. The CRC check is added to distinguish between upper-band payload and random bit-stream padding that can be added for probing available network bandwidth. At the receive side, a super-wideband payload is first given to the wideband decoder. The wideband decoder decodes as many parameters as required to uniquely reproduce the encoded wideband audio. The next byte in the payload should hold the value of LEN. This provides a sanity check that the decoding process has not failed. Thereafter, the receiver runs a CRC check over the upper-band payload and compares the results with the last 4 bytes in the packet. If the computed CRC and the last four bytes of the payload don't match, the remaining bits are assumed to be added for probing the network. Hence, the upper-band signal is replaced by zeros and combined with the wideband signal to generate the super-wideband signal. If the two CRCs match, then the upper-band payload is given to the upper-band decoder. Thereby, the output of the upper-band decoder is combined with the wide-band decoded audio to generate the super-wideband signal. It might be that for a given packet, the wideband decoder uses all the given payload. This can be the case when a super-wideband encoder is operating at low rates and has adjusted the effective bandwidth to wideband. In this case, the decoder inserts zeros as the reconstructed upper-band and combines both bands to reproduce the super-wideband signal.

The iSAC encoded upper-band speech data consists of parameters representing one frame of 30 ms speech. Depending on the target rate the upper-band encoder might choose to only encode the sub-band of 8 kHz to 12 kHz. This is signaled inband to the receiver.

Padding, which consists of randomly generated bits, may be added at the end of the payload in both wideband and superwideband modes. It can be used by the sender for bandwidth probing, and is always ignored by the receiver. In wideband mode, padding simply follows the payload, preceded by a length field.

Figure 5: Wideband payload format with padding. LEN is the length of the padding in bytes + 1: LEN = LEN_PAD + 1 In superwideband mode, the format of a packet with padding looks like the following.

| ]]> Figure 6: Super-Wideband payload format LEN is 1 + LEN_UB + 1 + LEN_PAD + 4, where LEN_UB is the length of the upper-band speech data in bytes, and LEN_PAD is the length of the padding in bytes. L2 is LEN_PAD + 1. The CRC check runs over the upper-band speech data, L2 and the padding.

More than one iSAC payload block MUST NOT be included in an RTP packet by a sender. Further, iSAC payload blocks MUST NOT be split between RTP packets.

This document defines the iSAC media type, and requests IANA to register it. audio isac None ibitrate: The parameter indicates the upper bound of the initial target bit rate the device would like to receive. For channel-adaptive mode, the target bit rate may vary with time; for channel-independent mode, the target bit rate will remain at that level unless instructed otherwise. An acceptable value for ibitrate is in the range of 20000 to 32000 (bits per second). maxbitrate: The parameter indicates the maximum bit rate the endpoint expects to receive. The recipient of this parameter SHOULD NOT transmit at a higher bit rate. This media format is framed and binary. See None RFC XXXX This media type is suitable for use in numerous applications needing to transport encoded voice or other audio. Some examples include Voice over IP, Streaming Media, Voice Messaging, and Conferencing. None COMMON iSAC is a proprietary speech and audio codec owned by Google. The codec operates on 30 or 60 ms speech frames at a sampling rate clock of 16 kHz or 32 kHz. Tina le Grand [tlegrand@google.com] This media type depends on RTP framing, and hence is only defined for transfer via RTP . Transport within other framing protocols is not defined at this time. IETF Audio/Video Transport working group delegated from the IESG. Note to the RFC Editor / IANA: Please replace "RFC XXXX" above with the number of this RFC when published, and remove this note.

The information carried in the media type specification has a specific mapping to fields in the Session Description Protocol (SDP) , which is commonly used to describe RTP sessions. When SDP is used to specify sessions employing the iSAC codec, the mapping is as follows: The media type ("audio") goes in SDP "m=" as the media name. The media subtype (payload format name) goes in SDP "a=rtpmap" as the encoding name. Any remaining parameters go in the SDP "a=fmtp" attribute by copying them directly from the media type string as a semicolon separated list of parameter=value pairs. The optional parameter ibitrate MUST NOT be higher than the parameter maxbitrate. The iSAC parameters in an SDP offer are completely independent from those in the SDP answer. For both ibitrate and maxbitrate it is legal for the answer to contain a value that is different than what is provided in an offer. The parameter may be present in the answer, even if absent in the offer. When conveying information by SDP, the encoding name SHALL be "isac" (the same as the media subtype).

The offer indicates that it wishes to receive a wideband bitstream with an initial target rate of 20000 bits per second. The remote party MAY change its initial target rate to the requested value.

The offer indicates that it wishes to receive a superwideband bitstream with an initial target rate of 20000 bits per second, and a maximum bit rate of 45000 bits per second. The remote party MAY change its initial target rate and SHOULD NOT transmit at a higher rate than 45000.

RTP packets using the payload format defined in this specification are subject to the general security considerations discussed in RFC 3550 section 8.1. As this format transports encoded speech, the main security issues include confidentiality and authentication of the speech itself. The payload format itself does not have any built-in security mechanisms. External mechanisms, such as SRTP , MAY be used.

This document was originally prepared using 2-Word-v2.0.template.dot. The present version is prepared using xml2rfc and xxe-xml2rfc.